핵심요약

• AI 중심 서버 랙의 전력 수요는 2030년까지 1MW에 이를 것으로 전망됩니다.
• 전통적인 서버 랙의 전력 소비는 15kW에서 80kW 수준입니다.
• GPU 기반 서버 랙은 최대 600kW까지 전력을 소비할 수 있습니다.
• AI 관련 서버 랙의 전력 수요는 기존의 컴퓨팅 용량보다 더 중요한 제약 요인으로 작용합니다.

도입

본 기사는 AI 시대의 성장을 뒷받침하는 핵심 제약 요인이 컴퓨팅 성능에서 전기 전력 공급으로 이동하고 있음을 제시합니다. 이는 투자자들이 데이터센터 및 AI 인프라 관련 기업을 평가할 때, 단순히 프로세싱 능력뿐만 아니라 안정적이고 폭발적인 전력 공급망의 변화를 반드시 고려해야 함을 의미합니다. 전력 인프라의 병목 현상이 새로운 시장 기회를 창출하고 있음을 이해하는 것이 중요합니다.

본문 1: 전력 수요의 질적 변화

전통적인 서버와 AI 중심 서버 랙 간의 전력 소비 패턴 차이는 현재 시장의 근본적인 변화를 보여줍니다. 기존 서버 랙은 CPU를 중심으로 작동하며 전력 소비가 15kW에서 80kW 사이로 비교적 안정적이고 예측 가능한 수준을 유지합니다. 반면, AI 중심의 Vera Rubin 랙은 GPU 기반으로 작동하며 최대 600kW에 달하는 전력을 요구합니다. 이러한 차이는 단순한 용량 증가가 아니라, 작업 수행 방식의 근본적인 변화를 반영합니다. 전통적인 시스템에서는 운영체제와 하드웨어가 수요 변화를 흡수했지만, AI 공장에서는 GPU가 벌크 동기식 패러다임 하에서 작동하며 전력 수요가 수 밀리초 내에 10%에서 100%로 급변하는 순간적인 전력 요구를 발생시킵니다. 이는 전력 공급의 안정성과 순간적인 반응성이 핵심적인 고려 사항이 되었음을 의미합니다.

본문 2: 인프라 투자와 시장의 반응

전력 수요의 변동성이 커짐에 따라, 산업은 새로운 문제 해결을 위한 전략을 모색하고 있습니다. 에너지 수요가 예측 가능한 안정적인 사용에서 폭발적이고 변동성이 큰 수요로 바뀌면서, 전기 그리드와 전력 공급 시스템의 개선에 대한 투자가 필수적인 과제가 되었습니다. 이러한 변화는 Bloom Energy, Infineon, GE Vernova, Delta Electronics와 같은 관련 기업들이 주가에서 트리플 자리 성장을 기록하는 배경이 됩니다. 이는 단순히 반도체나 컴퓨팅 기술에만 집중했던 과거와 달리, 에너지 변환 및 분배 기술을 제공하는 기업들이 새로운 성장 동력으로 부상했음을 시사합니다. 즉, AI 성장의 최종 병목은 하드웨어 자체에서 에너지 인프라로 이동하고 있으며, 이에 따라 에너지 관련 기업들의 가치가 재평가되고 있습니다.

본문 3: 장기적 전망과 리스크

AI 인프라가 2030년까지 1MW의 전력 수요를 예측하고 있다는 점은 장기적인 전력 인프라 구축의 필요성을 강조합니다. 이러한 폭발적인 수요를 충족시키기 위해서는 단순한 발전 용량 확대를 넘어, 전력의 효율적인 분배, 저장, 그리고 그리드 안정화 기술에 대한 혁신적인 접근이 요구됩니다. 만약 이러한 에너지 인프라 개선 속도가 AI 인프라 구축 속도를 따라가지 못한다면, 전력 부족으로 인한 병목 현상이 새로운 시장 리스크로 작용할 수 있습니다. 따라서 향후 투자 관점에서는 에너지 효율화 기술과 스마트 그리드 솔루션을 제공하는 기업들의 역할이 더욱 중요해질 것으로 전망됩니다. 이는 전력 공급망의 안정성이 AI 산업의 지속 가능한 성장을 결정하는 핵심 요소가 될 것임을 시사합니다.

결론

결론적으로, AI 시대의 성장은 컴퓨팅 능력뿐만 아니라 전기 전력 인프라의 제약을 받게 되었습니다. 전력 수요의 변동성과 폭발적인 증가에 대응하기 위해 에너지 솔루션 기업들의 역할이 더욱 중요해질 것입니다. 투자자들은 AI 인프라의 성장을 평가할 때, 하드웨어 효율성과 전력 공급망의 안정성을 통합적으로 분석해야 할 것입니다. 향후 에너지 인프라 기술의 발전 속도가 시장의 기대치를 충족시키는지 지속적으로 관찰해야 할 것입니다.

---

원문 링크: https://247wallst.com/investing/2026/07/07/musashi-seimitsu-might-be-the-missing-piece-in-power-distribution-for-datacenters/?.tsrc=rss

Musashi Seimitsu might be the missing piece in power distribution for datacenters

One of the sectors with highest growth in the stock market since 2025, has been electrical power related stocks. Names like Bloom Energy, Infineon, GE Vernova, and Delta Electronics have posted triple-digit gains . The fact that developers no longer size projects by compute capacity. They measure them in electrical power. Power, not processing, has become the real constraint.

Traditional servers in data centers essentially execute several tasks running in parallel. While some applications are running, others remain idle. The effect translates into a nominal, stable and predictable power usage. AI factories are the current paradigm for delivering AI inference at scale. They pressure the electrical grid at unprecedented rates. The difference with traditional server racks is in the amount of electricity and the instantaneous power variation.

Historically, the server racks use just CPUs for processing, and their power consumption varies between 15 kW and 80 kW. On the other hand, GPU-based Vera Rubin racks can reach up to 600 kW. Company reports such as Infineon’s expect the power needs in AI-focused server racks to surge up to 1 MW by 2030 .

Not only does power usage differ between traditional and AI focused server racks, the instantaneous power needs also change. In traditional servers, power draw stays mostly constant, and the operating system and server hardware absorb any changes in demand . In the AI factories, GPUs work in lockstep under the bulk synchronous paradigm. That translates into a unison execution of tasks in which power demand can go from 10 % to 100% in a couple of milliseconds.

‘This time is different’, is something we have all heard over the news regarding the stock market and the AI capex boom. At least for the electric grid, it seems like the change is real. As the energy demand changes from just predictable power usage, to a savage, explosive and volatile demand, the industry is developing different strategies to mitigate the new problems.

In a recent study with the name ‘ Power Stabilization for AI Training Datacenters ‘, researchers from OpenAI, Microsoft and NVIDIA published the power requirements for the AI Factories.

The main risks identified by the researchers arise from the violent swings in the power demand. The sudden changes induced by the synchronized computing loads, force tens, if not hundreds, of megawatts of power to be either required or be ready to be used in the grid. The problem is that the generation equipment is relatively slow and can’t react to the abrupt changes in power. When the load drops, the generation equipment must dissipate or absorb the excess power, which places mechanical stress on the machinery . On the other hand, if the grid can’t meet the power demand, the data center sits underused. To operate optimally and protect the data center’s equipment, operators must maximize instantaneous power efficiency.

The study advocates for cross-industry co-design through 3 strategies to maximize power efficiency and stabilize the power use. Software-based approaches are the easiest to deploy, using firmware and application software to smooth out violent power swings. During low-activity periods, the system injects controlled filler workloads to keep power draw steady. Examples include matrix multiplication sequences, GPU context sharing, and idle jobs. The approach is flexible and requires no hardware changes. It is the least efficient of the methods since it might impose performance limitations on the AI workloads, and the need for careful calibration of jobs, changes in the OS, and continuous monitoring.

Not only that, but the approach has limited effectiveness, as not all the power spikes can be filtered using this method. Consequences of this approach expand to the memory and compute resources needed just for monitoring and stabilization of the power usage. This translates to less compute resources for the business logic and overall energy burned. Finally, the coupling between application and infrastructure is far from ideal from a software architecture standpoint.

Second, the paper introduces the hardware level changes at server compute tray level. At this level power optimization approaches rely heavily on NVIDIA’s firmware and changes in the hardware introduced in the server blades. The firmware based approach focuses on limitations on the ramping rate for power usage. The feature enforces minimum power thresholds and regulates power fluctuations. Finally, the stop delay approach forces a minimum power usage threshold before the GPU finally ramps down. Similar to the software based solution, the wasted energy is the main downside of the hardware based approaches.

Third, the best-case solution to the training power-stabilization challenge is an energy-storage solution. It shall include enough capacitance to support the workload and handle the sudden rise/drop needs in power and can quickly switch between charge and discharge. It could charge during low-power phases and discharge during intense compute usage. The proximity with the rest of the compute modules makes rack-level storage the best option. The study highlights batteries and supercapacitors as the optimal approach to tackle the GPU induced power spikes.

The industry is working in parallel to match the power requirements, both at hardware and software level of present and future data centers. To illustrate, NVIDIA recently proposed a completely new power distribution architecture. At datacenter level, NVIDIA’s 800 VDC Architecture development path is described as the natural path to keep pushing the compute boundaries with more power efficiency.

Not only that, At server rack level electrical changes were introduced in Blackwell , the last generation of NVIDIA’s compute modules. GB300 NVL72, a Blackwell based server rack, designed to integrate hardware and software to smooth power spikes and reduce the peak demand up to 30%. In addition, NVIDIA’s last generation, VR200 NVL72 Vera Rubin based racks will increase the power consumption compared to GB300. This architectural change is expected to increase the number of passive capacitor components (MLCC) by 182% to respond to the high-frequency power transients and provide local energy storage around the computing units. In fact, Morgan Stanley analysis cited in recent coverage estimates the MLCC content in Vera Rubin VR200 at about $4320, compared to $1530 from Blackwell GB300 systems.

The focus of energy reliability extends to suppliers such as Vertiv ( $VRT ). The Battery Energy Storage System, DynaFlex BESS is used to set microgirds that use different power sources. The system guarantees Site-level energy assets are used to manage power demand, act as reliable backup during outages and constrain energy supply conditions and operational flexibility. The company also offers Battery based UPS systems such as Vertiv™ Liebert® EXL S1 UPS , designed to maintain uninterrupted operations of compute equipment with capacities up to 1200 kW.

Although these solutions act to stabilize the power supply and reliability of the grid, both Dynaflex BESS and Liebert EXL S1 rely on electrochemical batteries as the core energy buffer. The EXL S1 platform, for instance, handles 125% overload for 2 minutes, or 150% for just 15 seconds. Vertiv built the UPS for sustained, near-nominal power draw .

As previously mentioned, GPU clusters can swing from idle to 100% of load in milliseconds, and this is where batteries struggle to handle the quick load changes. Moreover, every cycle accelerates the battery degradation. That is why Vertiv developed features like Input Power Smoothing . The system uses its batteries as a short-term buffer to smooth input and output power peaks. This reduces stress on utility sources and the wider grid. The method works, but it carries several disadvantages. Those drawbacks tend to limit the system’s performance.

The most important is perhaps that the battery behaves asymmetric in load and discharge conditions. In addition, the rate at which batteries can charge and discharge is also limited to a certain threshold, which leads to oversizing the batteries for the system. To stabilize the load, batteries must be properly sized and be capable of sustaining very high charge rates. High C-rate batteries let the UPS absorb and inject large amounts of power within a fraction of a second. This compensates for sudden AI workload swings and keeps input demand from the generator steady.

An undersized system limits the performance and compromises long-term reliability. From an energy storage perspective, when the battery recharge rate is too low, the UPS reduces its ability to sustain input power smoothing (IPS), allowing load fluctuations to propagate upstream and placing additional stress on generators and the electrical grid.

Vertiv’s approach can stabilize load transients to a certain extent, but the batteries’ chemistry ultimately limits it. As previously mentioned, the alternatives rely on physical rather than chemical energy storage. Supercapacitors can fill the role in high power quick transient mitigation and several products in this category already exist. To illustrate, Flex’s ( $ FLEX ) CESS is a capacitive based energy solution for AI datacenters, built around Musashi Energy Solutions’ Hybrid SuperCapacitor (HSC) Cells. Musashi Energy Solutions also offers their own equipment based on supercapacitors. For instance the ESS400 Energy Storage System can handle AI workloads with power capacities up to 400 KW per rack which can be combined to support operations of up to 2 MW.

Super capacitor cells can act on the same physical timescale as the load itself, thus reacting in milliseconds. Unlike battery chemistry, supercapacitors charge and discharge symmetrically. That symmetry removes the need to oversize the cells. In addition, Musashi’s cells reach peak discharge currents of up to 1,200 A, withstand up to 1 million charge-discharge cycles, and carry no thermal-runaway risk — unlike lithium-ion batteries.

Furthermore, Musashi’s hybrid supercapacitor cells combine high power density with a prismatic format. That design dissipates heat and packs more efficiently than competitors’ cylindrical cells. Compared to batteries, HSC cells tolerate one to two orders of magnitude more cycling than even high C-rate lithium-ion. They also charge and discharge symmetrically. Lithium’s asymmetry forces Vertiv’s IPS algorithm to oversize battery capacity just to keep recharge ahead of discharge.

Architecturally, the two approaches solve different problems: Vertiv’s UPS targets outages, while Musashi’s HSC-based buffering isolates load transients. UPS and HSC can be employed in synergy to reduce stress on the electric grid.

Musashi’s advantage is clear: prismatic design of supercapacitors and the highest publicly verifiable power density for supercapacitors. This approach makes it ideal to match the industry needs in power transients. However, the company is not the only nor the largest supplier in the supercapacitor industry. Companies like Eaton ( $ETN ), Kyocera ( 6971.T ), Panasonic ( 6752.T ) and Murata ( 6981.T ), to name a few, also produce supercapacitors, are more diversified, and have larger market caps than Musashi.

On top of that, Musashi appears to be in a weak financial position. The parent company, Musashi Seimitsu Industry Co. ( 7220.T ), reported full-year revenue mostly flat at ¥347B ~$2.3B), while net income collapsed 84% YoY to just ¥1.26B, cutting net margin to 0.4%. The stock has also crashed abruptly, losing roughly 62% of its value from ¥10,550 to ¥4,035 in June 2026.

Power generation remains a priority for data center developers, and Musashi’s technology might be a missing piece. Still, its financials suggest this thesis remains early-stage.

Source: https://247wallst.com/investing/2026/07/07/musashi-seimitsu-might-be-the-missing-piece-in-power-distribution-for-datacenters/?.tsrc=rss